Key ideas from Richard Sutton

• "Large language models learn from what people say, not from experience. They mimic rather than understand the world" - Richard Sutton
• Supervised learning doesn't occur in nature - animals learn through trial-and-error and prediction, not from examples of desired behavior
• "Intelligence is the computational part of the ability to achieve goals" - Sutton citing John McCarthy's definition
• Gradient descent alone won't produce good generalization - it finds solutions but doesn't inherently transfer well to new states
• The bitter lesson may repeat: methods using human knowledge in LLMs could be superseded by pure experiential learning systems
• "We're entering the age of design - transitioning from replicated intelligence to designed intelligence we actually understand" - Sutton
• Continual learning agents need four components: policy, value function, state representation, and transition model of the world
• Digital intelligence succession is inevitable based on: lack of unified control, eventual understanding of intelligence, reaching superintelligence, and intelligent systems gaining resources